healthtransformationlab/perio-bench-notes

PerioBench notes: synthetic dental notes labelled with periodontal status

3,484 synthetic dental chart notes to train on and 240 to test on, each labelled with the patient's periodontal status under the 2017 AAP/EFP classification: healthy, gingivitis, or periodontitis stage I to IV, with grade and extent. Built to practise clinical NLP and phenotyping from dental records without touching real patients. Part of PerioBench from the Health Transformation Lab.

Quick start

from datasets import load_dataset

notes = load_dataset("healthtransformationlab/perio-bench-notes")   # train: 3,484 notes, test: 240
print(notes["train"][0]["note_text"], "->", notes["train"][0]["label"])

Or with pandas only:

import pandas as pd

train = pd.read_csv("hf://datasets/healthtransformationlab/perio-bench-notes/data/train.csv")
test = pd.read_csv("hf://datasets/healthtransformationlab/perio-bench-notes/data/test.csv")

Try it in your browser: The quickstart notebook (also in this repository) trains a simple baseline in about a minute and scores it on the test notes beside the published leaderboard. The same data and notebook are on Kaggle: kaggle.com/datasets/jcblearning/perio-bench-notes.

What's inside

Config	What it is
notes (default)	train and test splits: the note, its label (healthy, gingivitis, periodontitis_stage_1 to 4), grade, extent, what the clinician wrote (stated_status), and the X-ray bone-loss reading
review_key	The 240 test notes with every other part of the record: exam form, procedure and diagnosis codes, probing chart, what each says, and a difficulty level (L1 to L3)
visits, visits_truth	Every visit's structured record and its true status, for phenotyping from structured data
clinics	The 8 synthetic clinics

Score on test; don't train on it. The raw/ folder holds the original files, as on Kaggle.

Everything here is synthetic. No patient, no PHI, no real clinic. The generator wrote fake identifiers into the notes and a de-identification step replaced them with placeholders such as [NAME] and [DATE]. This is a personal learning project, not affiliated with any employer, institution or company, and not for clinical use: nothing here is evidence about real patients, products or populations.

How it was made

github.com/healthtransformationlab/health-transformation-lab-nlp, python/scripts/10_build_study.py (seed 42, 1,200 people, 8 clinics). Each visit gets a true periodontal chart; the status follows the AAP/EFP staging and grading worksheet (Tonetti, Greenwell & Kornman, J Periodontol 2018). The note is written from the chart in one of several clinician styles, with realistic noise: the written diagnosis is sometimes a step off the chart, text is copied forward from earlier visits, and older terms ("moderate chronic periodontitis") appear.

Licence and citation

CC BY 4.0. Please cite: Bowyer J., health-transformation-lab-nlp: synthetic dental notes, OMOP and periodontal status benchmark, https://github.com/healthtransformationlab/health-transformation-lab-nlp.

Contents

File	Size
data/test.csv	240 rows
data/train.csv	3,484 rows
perio_bench_quickstart.ipynb	notebook
raw/clinic_profiles.csv	8 rows
raw/review_key.csv	240 rows
raw/review_tasks.json	240 items
raw/train.csv	3,484 rows
raw/visits_structured.csv	4,343 rows
raw/visits_truth.csv	4,343 rows

Published from commit 8a3e5139880d387aff398c70aa91e72beeb41a05 of the repository.